Start with the number that should give anyone pause. The biometric identity firm iProov showed 2,000 people in the UK and US a mix of real and AI-made images and video, and asked them to spot the fakes. Just 0.1% got every one right — one person in a thousand — and these people knew what they were looking for, having been told to watch for deepfakes before they started.
One thing to keep in mind throughout. iProov sells deepfake-detection technology. This is a company-commissioned survey, not peer-reviewed research, and the firm has a commercial reason to want a scary result. That does not make the number wrong. It does mean the framing deserves a careful eye, which is the point of this piece.
The number that should give us pause
The 0.1% counts only the people who got every single image and clip right. Miss one, and you are out of that group. It is a tough test, and the headline gap partly reflects how strict it is. Still, the direction is hard to ignore. “Just 0.1% of people could accurately identify the deepfakes,” said iProov founder and chief executive Andrew Bud, “underlining how vulnerable both organizations and consumers are to the threat of identity fraud in the age of deepfakes.”
The broader research is not quite as bleak as that headline suggests. A 2024 review by Diel and colleagues pulled together 56 studies and 86,155 participants in the journal Computers in Human Behavior Reports. It did not find that people were helpless. The authors report that “total deepfake detection accuracy was 55.54%.”
That is a different kind of warning. Not “almost nobody can ever spot a fake”, but “ordinary human judgement is a weak and inconsistent safeguard.” Across tens of thousands of people, telling real from fake came out much closer to a coin toss than most of us would like to believe.
Why our confidence stays fixed regardless of whether we’re right
Here is the part that unsettles more than the accuracy figure. In the iProov study, over 60% of participants stayed confident in their ability to spot deepfakes, and that confidence held whether or not their answers had been right. The feeling of “I can tell” survived the evidence that, mostly, they could not.
Again, other research points in the same direction. A separate study of people judging fake audio and video found average accuracy of 65.64%. But confidence ran higher still, at 77.60%. People were not useless at the task. They were just more confident than their performance justified.
That gap matters. Low accuracy can be trained against; low accuracy paired with high confidence is harder, because the people getting it wrong do not always know they are. Anyone who has confidently told a friend a viral clip was obviously fake, without ever checking, has felt the small version of this.
Video is harder than images, and the gap matters
The iProov data also points to where the weakness sits. Participants were 36% less likely to correctly identify a fake video than a fake image. The moving, talking fake, the format closest to a video call or a verification check, was the one people read worst.
The review tells a more mixed story. Across its data, there was no clean “video is always hardest” rule. In fact, pooled accuracy ran a little higher for video and audio than for still images. That does not cancel out iProov’s finding. It simply means format effects depend on the study: what kind of fake was shown, how good it was, how long people had to look, and whether they were judging a still image, a recorded clip, or something closer to a live interaction.
So the narrower, more defensible read is this: video-based checks deserve particular care, especially when they are used in video calls or identity-verification settings. The evidence does not prove that people always fail with video. It does show that we should be very cautious about treating a visual judgement as enough.
What this means for systems built on human eyes
The conclusion the study reaches for is, conveniently, the one iProov is in business to sell. Edgar Whitley, a digital identity expert at the London School of Economics, is quoted in the release arguing that “organizations can no longer rely on human judgment to spot deepfakes and must look to alternative means of authenticating the users of their systems and services.” That is one expert’s view inside a vendor’s promotional release, and the “alternative means” points neatly at iProov’s own product. Read it as a direction of travel, not a settled verdict.
The underlying concern is real, though. Bud’s blunter line, that “criminals are exploiting consumers’ inability to distinguish real from fake imagery,” may overstate the cause and effect, but it names a genuine weak spot. Any identity process that still relies on a person visually judging whether a face “looks right” should treat that judgement as weak evidence, not proof.
There is a more hopeful thread, and it sits in the review rather than the vendor release. Diel and colleagues found that giving people feedback and software help pushed accuracy above chance. Human judgement is not fixed at coin-toss level — it is poor when left untrained and unsupported, which describes most people most of the time.
The confidence finding leaves one question that should keep system designers up at night. If the people who are wrong feel exactly as sure as the people who are right, confidence tells you nothing about accuracy. Any process that quietly trusts a human’s “I’m sure that’s real” is trusting a call the evidence says the human cannot reliably make.
Editorial process
Space Daily articles are produced with AI assistance and reviewed by editorial staff before publication. See our editorial standards and masthead.
